The video reveals a major leak from Google DeepMind showcasing Gemini 3, an advanced AI model capable of recreating entire operating systems in-browser, generating complex 3D visuals, and solving advanced problems, alongside Nano Banana 2, a powerful image generation AI, and a new autonomous web agent prototype. These innovations highlight significant leaps in AI creativity, technical skill, and practical applications, signaling an intense competitive race in AI development for the near future.
The video reveals a major leak from Google DeepMind showcasing three next-generation AI products: Gemini 3, Nano Banana 2, and a new agent prototype. Gemini 3 stands out as the most advanced model, demonstrating remarkable speed, intelligence, and competence. One of the most impressive feats shown is Gemini 3’s ability to recreate entire operating systems like Mac OS and Windows directly in the browser with functional apps such as file explorers, text editors, terminals, and games. These creations are achieved through clever JavaScript tricks and creative problem-solving by the model, highlighting its innovative capabilities.
The leak originates from Google’s internal process of developing AI models through “hero runs,” where the best ideas and experiments are combined into a new version. Gemini 3 is not a single model but a collection of multiple branches, with rumors suggesting around 13 variants being tested. These models are exposed to real users through AI Studio for feedback, which helps Google refine and select the best version for official release. Currently, users can access different versions of Gemini 3, including a powerful but unreleased variant called ECPT, which has demonstrated exceptional coding and creative abilities.
Gemini 3’s capabilities extend beyond coding and OS simulation. It can generate complex 3D visualizations, create interactive games with menus and animations, and produce detailed SVG art that can be manipulated. The model has also solved advanced mathematical problems previously thought unsolvable by language models, showcasing its deep reasoning skills. Despite some skepticism about certain examples, the overall performance of Gemini 3 is a significant leap forward in AI creativity, technical skill, and aesthetic quality.
Alongside Gemini 3, Google has developed Nano Banana 2, an upgraded image generation and editing AI that surpasses its predecessor in realism, instruction following, and 2D graphic creation. Although briefly leaked and quickly taken down, Nano Banana 2 impressed with its high-quality outputs, including 2K native resolution and 4K upscaling, transparency features, and improved understanding of complex visual concepts. It is expected to be released soon, possibly alongside Gemini 3, marking another major advancement in Google’s AI media capabilities.
Finally, the video touches on a new agent prototype from Google DeepMind designed to perform web-based tasks autonomously. Unlike OpenAI’s agents that allow user control, this agent operates independently to complete tasks such as visiting websites and handling design commands within Google Stitch. While still in the research preview phase, this agent shows promise for practical applications. The video concludes by anticipating a competitive end to the year in AI development, with OpenAI likely to respond with new releases, keeping the race for AI innovation intense and exciting.